Pattern Identification

How do I know whether a pattern in my findings is real?

Four checks, all of which can fail. Count the search space: how many comparisons or pairings were available, and how many would look patterned by chance. Screen for shared cause in the instrument: adjacency, shared batteries, shared routing, shared bases, shared method. Count the findings inside the pattern's scope that it does not cover, and report that number with it. Then state what observation would break the pattern and go and look for it in evidence you did not use to build it. A pattern that has passed all four is a claim. One that has passed none is an arrangement.

How many findings do I need before I can claim a pattern?

Roughly six to eight independent findings, and independence matters more than count. Below that, almost any arrangement will look patterned, because the number of possible arrangements is small and the eye completes them. Findings drawn from the same battery, the same routing branch or the same respondents are not independent; five of them may be one finding described five ways. Count independence after the instrument screen, not before.

Why do three questions correlate when the underlying attitudes do not?

Because they sat next to each other. Adjacent items prime one another, shared stems and scales invite a consistent response style, and a respondent who has just answered one question about a service answers the next three in its light. The diagnostic is that the items correlate with each other about as strongly as any of them correlates with anything outside the block. That is what shared measurement looks like, and it is not evidence that the attitudes cohere.

What is the difference between a pre-specified and an exploratory pattern?

A pre-specified pattern was predicted before the data was seen, then observed. An exploratory pattern was selected from among all the patterns the data could have shown, and the number of those is large. Both are legitimate outputs; they are not equally strong, and the difference has to be labelled. If the prediction was written after the search, it is not a prediction, and treating it as one is the research equivalent of moving the target after the shot.

What pattern types should I actually look for?

Seven. Co-occurrence (do these appear in the same units?), sequence (does one reliably precede the other?), threshold and non-linearity (is there a point where the relationship changes?), segment-specific effects, absence where something was expected, consistency across measures that share no method, and universal versus local. Sequence and absence are the two most often missed: a findings list has no time axis until you build one, and nothing can be absent unless someone wrote down what was expected.

Is a finding that appears in one market a pattern?

It is a pattern about that market, and the qualifier is part of the claim rather than a caveat attached to it. Write the scope inside the sentence, because qualifiers that live in footnotes are gone by the time the summary is written. Note also that "consistent across all markets" means something very different for two markets than for eleven, so state the number.

What do I do with findings that fit no pattern?

Keep them, list them in full, and say they were not forced into one. Unpatterned findings are not evidence of failed analysis, and some of them will be the most important results in the study. The alternative, stretching a claim until it covers everything, produces a pattern that explains nothing and that will not survive its first serious challenge.

Should I explain the pattern at the same time?

No, and the reason is practical. The moment a candidate acquires a "because", it becomes much harder to abandon, and abandoning candidates is the entire purpose of this stage. Establish that the structure exists, define its edges and its exceptions, then hand it to insight development to explain.

What if no pattern survives testing?

Report that. Give the findings individually, state the search space, list the rejected candidates and why each failed, and say that the evidence does not support a structural claim. This is a normal result in studies with few independent measures. A report with eleven findings and no pattern, honestly labelled, protects every conclusion downstream; a manufactured spine breaks all of them at once.

How do I check an AI-generated theme?

Take it as a single candidate rather than a conclusion. Ask which respondents or cases each grouped finding actually spans, because language models group by similarity of wording and level of abstraction, not by shared units in the data. Then run the instrument screen, count the findings in scope the theme fails to cover, and ask what would break it. Most proposed themes fail at the exception count, and producing that count is a fast way to redirect a report before it is built.